How one Virginia county’s AI platform protects sensitive data
John M. Chase via Getty Images
By
Chris Teale
Prince William County uses Liminal to prevent employees from inputting sensitive data into generative AI in what a spokesperson described as a “filtration level.”
Governments of all sizes have been rolling out generative artificial intelligence tools to their employees, and one county looks to have found a solution to prevent staff from using sensitive data on the platforms.
Prince William County, Virginia uses a platform that takes a generative AI prompt from a member of staff, then either redacts or removes any sensitive data while meeting all applicable security standards. After sending the secure prompt to a generative AI platform and receiving an answer, it rehydrates it with sensitive data.
Dr. Alicia Hart, deputy county executive for government operations, performance and innovation, said this method ensures the context of a prompt remains, but county staff and leadership can be sure it complies with various data protection laws and requirements.
“One of the things that my team continued to come to me about were concerns regarding how widely used [large language models] were within the organization, but we didn't necessarily have guardrails in place to help support and protect sensitive data,” Hart said during a panel discussion at the Ivanti Public Sector Summit near Washington, D.C. earlier this month.
For example, Hart said, an employee in the county’s Department of Social Services might use the platform to remove any sensitive and personally identifiable information from a report before asking a generative AI platform to summarize it. That platform would then identify sensitive data, redact or remove it, send the prompt out, then put the sensitive data back in for the employee once the response comes back.
“That then allows the intent and the context of the prompt to remain, while we are also remaining within compliance,” Hart said.
In an email after the event, a county spokesperson said Prince William “wanted to ensure that the information being put into prompts met all applicable security standards,” and “evaluated tools that could protect that data at the point of use.” The spokesperson said the county chose Liminal, a platform that describes itself as providing AI “total oversight and ensuring enterprise‑wide compliance” to fulfil that role after a pilot phase and a full infrastructure review.
Hart declined to answer further off-stage questions on the initiative from a reporter after the event, and a county spokesperson did not make any member of government staff available for an interview. Liminal did not respond to requests for comment.
“That [review] included working sessions with Liminal's C-suite to understand the data flow end to end: what the system recognizes, how it redacts versus masks depending on the use case, and how it behaves across different scenarios before any broader rollout,” the spokesperson continued in an email.
On stage, Hart described the rollout as a “soft launch,” which the spokesperson said in an email was a “pilot phase with 150 users across several departments.” Then, using its findings from that pilot, the county’s Department of Information Technology presented it to executive management and agency leadership for their approval. The IT department then shared it with employees through its Communications and Engagement Office. As of now, the spokesperson said, 20 departments are using the tool.
“The department users have really been the greatest advocates for the product itself,” Hart said on stage at the Ivanti summit. “Now they are using it, they are encouraging their colleagues to use it, and eventually we will get to the point where it is an enterprise-wide requirement. But we're doing so in a way that definitely involves our end users and stakeholders and having them at the table.”
The effort comes as Prince William County has worked to embrace AI throughout its government operations. Last week, the county announced that its PWC311 service — known as Will and powered by AI — recorded nearly 90,000 interactions with residents across its web, mobile and AI-powered channels and nearly 40,000 questions to help provide customer support and direct them to the relevant department for assistance. Hart said on stage more is to come.
“We are definitely moving forward with increasing innovation within the organization, but also wanting to find ways to support our staff members and our employees in making work more efficient,” she said. “When we launched this soft launch, we had a number of agencies say, ‘Yes, we would like to use this, we would like to sign up.’ Our philosophy is not necessarily forcing it upon the departments, but rather gaining that buy-in.”